Training settings
Last updated: October 15, 2026
Every setting on the training form, by the label it has on screen: what it does, its range, and when to move it. Use it when you train in Customize or adjust a run after reviewing a finetune.
Defaults depend on the base model and are listed in Trainable models. Why the settings interact the way they do is in How training works. Fix the dataset and captions before you change any of these.
Choosing how to train
Automatic
Layer sets everything up from the set's images: it picks the base model, uses that model's default settings, writes any missing captions, and gives the finetune a keyword. It trains an image finetune from the images and, when the set has enough complete before-and-after pairs, an edit finetune as well. Use it for a first version. If the set doesn't have enough images yet, the button says Add more images to train.
Customize
You choose the base model and the settings yourself. The trigger word and the trainer's settings sit behind Show advanced settings; without opening it, the run uses the model's defaults.
Base model
The model the finetune trains from, and the only model it works on afterwards. Pick the one you will generate with.
Trainer
Which engine trains the finetune. Most models offer only fal. FLUX.1 [dev] also offers Kohya, an open-source trainer; Qwen-Image and Qwen-Image 2512 also offer AI Toolkit, which has extra settings (below). Switching trainer resets the settings below it to that trainer's own defaults. Kohya and AI Toolkit cost more per training step than fal.
Dataset and captions
References
The set's files, under References (or Training pairs for an edit model). Only image files train an image or edit finetune. The trainer works at a fixed resolution, so use the full-size source of each image; small, blurry or upscaled images teach blur. An image finetune needs a minimum number of images, and the button tells you how many more to add.
Captions
The text under each image when an image finetune is open (Add a caption… on an empty one). Images you don't caption are captioned automatically when training starts; captions you wrote are never replaced. Describe what varies between images and leave out what should always be there. See Captions and trigger words.
Describe the edit
The editing caption under each before-and-after pair, for edit finetunes. Every pair you train needs one. Apply to all pairs writes one description to every pair, either replacing what is there or filling only the empty ones.
Trigger word
A keyword the finetune ties to the concept, the first field under Show advanced settings. Layer adds it to the training text and to your prompt when the set is attached, so you don't type it in captions. Pick a word with no ordinary meaning. Character sets anchor on the set's name instead.
Training settings
Epochs
How many times the model sees the whole set. Together with the number of images, epochs set the training steps, which set the cost and the time. The slider runs from a floor that still trains a usable finetune to a maximum the trainer allows; a very large set may train fewer epochs than you set, to stay under the trainer's step limit.
Raise it when the finetune is weak: the concept only shows at high weight, or details drift. Lower it when results copy the training images, ignore the prompt, or look burnt, and when a run timed out (Adjust settings on a timed-out run lowers it for you).
Learning rate
How large each adjustment is. It pairs with epochs: higher epoch counts usually want a lower learning rate.
Raise it when the finetune is weak and you don't want a longer run. Lower it when results overfit or colours go harsh, and whenever you raise epochs a lot. Change it in small steps; it has more effect than its small numbers suggest.
LoRA rank
The LoRA's capacity: how much detail it can hold. A higher rank captures finer detail and more variety, and makes a bigger weights file; it also memorises a small set more readily. A lower rank holds the broad look and generalises well.
Raise it when a large or detailed set loses fine detail (texture, small costume parts, intricate line work). Lower it when a small set, or a single character, starts copying its training images.
Preview prompts
Sample prompts rendered during training, at each saved checkpoint, so you can see how the finetune is turning out. They don't change what is trained. Preview images are charged as generations.
Include one prompt close to the training images and one that asks for something the set doesn't show, so the previews tell you whether the finetune generalises.
Generation defaults
Training doesn't change the generation form: steps, guidance and the other generation settings start from the base model's own defaults whenever you generate with the finetune. Two things from the set reach every generation that uses it.
Trigger word at generation
The trigger word (or a character set's name) is added to your prompt when the set is attached. You don't need to type it.
Always applied
Text from the set's Always applied prompt joins every prompt that uses the set. It is not part of training, so you can change it at any time without retraining. Use it for framing or background notes you want on every image.
The finetune's strength is set per generation with its weight control. See Generate with a trained reference set.
AI Toolkit settings
Shown under Show advanced settings when Trainer is AI Toolkit, grouped under headings such as Network, Dataset, Checkpoints and Sampling.
Network alpha
Scales how strongly the LoRA's learned changes apply, relative to its rank. From 1 to 256; left empty, it equals the LoRA rank.
Raise it when the finetune learns too slowly at a learning rate you want to keep. Lower it when training overshoots; lowering alpha softens each update much like a lower learning rate. Most runs leave it at the rank.
Optimizer
The method that turns each step's error into an adjustment: adamw8bit (default), adamw or adafactor. adamw8bit is a memory-saving form of adamw and learns much the same; adafactor adapts its own step size. Keep the default unless you are comparing runs on purpose.
Noise scheduler
How noise is added and removed during training. flowmatch is the only option, matched to the Qwen-Image models.
Gradient accumulation
How many images the trainer looks at before it makes one adjustment. From 1 to 16; default 1.
Raise it when learning is erratic on a varied set: each adjustment then averages over several images, which is steadier but slower to learn. Lower it when the finetune learns too slowly. Most runs leave it at 1.
Max saves to keep
How many checkpoints the trainer keeps during the run. From 1 to 100; default 100. It doesn't change what is trained.
Caption dropout rate
The share of steps on which an image trains without its caption, so the concept also holds when a prompt says little. From 0 to 1; left empty, no captions are dropped. Character sets start at 0.3.
Raise it when the concept only shows up with prompts that resemble the captions, or a character drifts unless described. Lower it when prompts stop steering pose, setting or framing.
Sample every (steps)
How often the trainer renders preview samples, in training steps. Default 250.
Sample steps
Denoising steps for each preview sample. From 1 to 50; default 25. Fewer is faster and rougher, and doesn't affect the finetune.
Sample guidance scale
How closely preview samples follow their prompt. From 1 to 10; default 3. It affects only the previews.
Voice settings
Method
Clone copies a voice from reference clips in the set. Design creates a voice from a written description. Which methods appear depends on the model.
Voice description
For Design: the voice to create, in words: tone, age, accent, delivery. On ElevenLabs TTS V3 it takes 20 to 1000 characters.
Clip to clone
For Clone on Qwen 3 TTS or Seed Audio 1.0, which clone from one clip, when the set has several: which clip to use. On Qwen 3 TTS a clip can be up to 300 seconds long. ElevenLabs TTS V3 clones from several clips at once, so it doesn't ask.
Reference clip
For Design on ElevenLabs TTS V3, when the set has several clips: the one clip that guides the designed voice.
Reference transcript
For Clone on Qwen 3 TTS: the exact words spoken in your clip. Optional, and improves the clone when supplied.
Auto script
For Design on Qwen 3 TTS: Layer supplies a script for the designed voice to read. Turn it off to write your own in Script to speak.
Script to speak
The words the designed voice reads aloud, when Auto script is off.
Loudness
For Design on ElevenLabs TTS V3: how loud the designed voice is, from -1 to 1. Raise it when the voice sits too quietly under music or effects; lower it when it sounds pushed.
Guidance scale
For Design on ElevenLabs TTS V3: how strictly the voice follows your description, from 0 to 100. Raise it when the voice ignores parts of the description; lower it when it sounds exaggerated or unnatural.
Prompt strength
For Design on ElevenLabs TTS V3 with a reference clip: how strongly the clip steers the voice, from 0 to 1. Raise it when the voice should sound more like the clip; lower it when the description should lead.
Motion settings
Method
Text generates a motion from a description. Video extracts the motion from exactly one reference video in the set.
Motion description
For Text: the action, its style and its pacing, from 1 to 1000 characters. Describe one motion, as you would brief an animator: "a heavy two-handed overhead sword swing, slow wind-up, fast follow-through".
Model version
The Uthana Animate generator version. Text offers text-to-motion-3.0 (default) and text-to-motion; Video offers video-to-motion-2.1 (default) and video-to-motion-2.0. Keep the default unless an earlier version suited a motion better.
Image style settings
Match
Precise: generations follow your references closely. Flexible: generations keep the feel of your references but stray further for prompts they don't cover. Choose Precise for a strict brand look, Flexible when you will prompt subjects far from the references.
Reference weights
How much each reference counts towards the style, set with a slider under each image, from 0.1 to 2. Left even, all count equally. A run uses between 1 and 10 references; beyond that, only the first are used and the form says how many are left out.
Raise a weight when one reference best captures the look. Lower it when an image is in the set for its subject rather than its style.